Systematic Literature Review of Continuous Validation and Improvement Methods for AI Systems

Abstract

This systematic literature review aims to systematically analyze current methods of continuous validation and improvement in AI systems. It investigates techniques, application domains, scope of implementation, and real-world challenges to identify prevailing trends, limitations, and opportunities for enhancing adaptive, reliable, and ethically responsible AI deployment This systematic review employed a structured search across IEEE Xplore, Scopus, Web of Science, and ACM Digital Library (2018-2025), using defined keywords. Studies were screened through title, abstract, and full-text review, applying inclusion/exclusion criteria. Quality was assessed using context, design, validity, rigor, and relevance dimensions. A total of 51 studies were analyzed. Cross-validation (25.49%) and online retraining (23.53%) were the most used validation methods. Improvement efforts centered on iterative model refinement (21.57%) and integration with feedback and workflows (15.69% each). Most studies focused on smart manufacturing and robotics (each 27.45%), with healthcare (13.73%) and environmental systems (11.76%) trailing behind. Industry-focused deployments dominated (19.61%), while only 9.80% addressed cross-regional implementations. Key challenges included model drift (13.73%), generalizability issues (11.76%), and ethical concerns (11.76%). Real-time feedback mechanisms, regulatory alignment, and interpretability remained under-addressed, signaling critical gaps in sustainable and trustworthy AI development. While technical validation and improvement strategies are maturing, gaps persist in real-world adaptability, ethical integration, and socio-technical feedback loops. Bridging these gaps will require collaborative, context-aware, and regulatory-informed AI systems capable of maintaining performance and trust across diverse, evolving environments.